MI nodes

A pack of seven ComfyUI custom nodes for everyday Stable Diffusion work: stack LoRAs with remembered (and Civitai-searchable) trigger words, pick embeddings from a dropdown, rate your diffusion models, compare three images side by side, save only the results you like, and get a sound when a long run reaches a point.

All nodes live under one menu category, MI nodes, and show the same MI nodes badge.

Node	What it does
MI Lora	Several LoRAs in one node: on/off per row, strength, remembered trigger words and a Civitai lookup
MI Unet Rate	UNET / diffusion-model loader where you rate (1-3 stars) and comment every model
MI Embedding	Embedding picker with strength, on/off switch and several rows
MI Compare 3 Images	Compare up to three images inside the node: switch, wipe, save
MI Save Image	Preview in temp, copy to output only when you press SAVE
MI Reroute Sound	A minimal reroute that plays a WAV sound when the flow passes through it
MI Bounding Boxes	Draw labelled boxes on a canvas inside the node (classic LiteGraph) for MOSS-Image and Ideogram

Installation

With git (recommended, always the latest version). Clone into your ComfyUI custom_nodes folder:

cd ComfyUI/custom_nodes
git clone https://github.com/ThinkingBells/MI_nodes.git

With ComfyUI-Manager or comfy-cli. Search for MI nodes in the Manager, or run comfy node install mi-nodes. The registry publishes each version after a review, so it can be a little behind this repository.

Then restart ComfyUI and reload the browser tab (Ctrl+F5). No extra Python dependencies are needed.

Keep the folder name MI_nodes. It must not contain spaces: ComfyUI cannot serve the interface files of a folder with a space in its name.

If you previously installed the standalone packages (MI_lora, MI_embedding, MI_unet_rate, MiCompare_3_images, MIsave_image), remove them first or the nodes will appear twice. The internal node names are unchanged, so existing workflows keep working.

To update: git pull inside the folder and restart. To find a node: double-click the canvas and type its name, or browse the MI nodes category.

Compatibility: developed and tested with ComfyUI 0.42 and frontend 1.55 (classic canvas), Python 3.10+.

---

MI Lora

Loads several LoRAs from one node and remembers the trigger words of each one. Write the trigger words once: they are saved per LoRA, filled in automatically the next time you pick it, and sent as text to your prompt. Not sure what a LoRA's trigger words are? Look them up on Civitai from the node.

Why: stacking LoRAs usually means a chain of loader nodes plus a notes file to remember which words activate each one. MI Lora puts the LoRAs, their strengths, their trigger words and an on/off switch in one compact node.

Features

• Several LoRAs in one node: + / - buttons add or remove rows (up to 8).
• On / off per LoRA: a row that is off is skipped (the LoRA is not applied and its trigger words are not sent). Its fields are greyed out and the saved words are kept.
• One strength per LoRA: from -10 to 10, applied to both model and clip.
• Trigger words per LoRA: saved for that LoRA and filled in whenever you pick it again, in any node or workflow.
• Civitai lookup: press search triggers on Civitai at the bottom of the node. Every active LoRA is looked up by its file hash and a framed box lists the trigger words found, one numbered result per row. Click a result to put its words in that row's triggers field. Long lists wrap onto several lines. Results without trigger words (or not found) are shown in italics and are not clickable.
• Sent as text: the text output is your incoming prompt followed by the trigger words of every active LoRA.
• Subgraph friendly: every row uses standard widgets.

Usage

1. Connect model and clip, and optionally a STRING prompt to text.
2. Pick a LoRA, set its strength and write its trigger words (comma separated), or fetch them from Civitai.
3. Connect model and clip to your sampler and text encoder, and text to the text input of a CLIPTextEncode (convert the widget to an input first).

Text in	LoRA rows	text output
1girl,	add_detail (on), Cute Faces (on)	1girl, detail, cute faces
1girl,	add_detail (on), Cute Faces (off)	1girl, detail
nothing connected	add_detail (on)	detail

Rows set to (none) are ignored.

Civitai lookup and privacy. Nothing is sent until you press the search line. Then only the SHA256 hash of each active LoRA file is sent to civitai.com, never the file itself. Hashes are cached in data/hashes.json, so each file is hashed only once (the first search on a large LoRA can take a few seconds). A LoRA that is not on Civitai, or that lists no trigger words, is reported as such.

Good to know: trigger words are tied to the LoRA path and file name inside the loras folder; renaming a file makes it a new LoRA. The LoRA list is read at startup, so restart ComfyUI after adding files.

---

MI Unet Rate

A UNET / diffusion-model loader that lets you rate and comment every model right where you pick it. File names alone are rarely enough to remember which of your dozens of checkpoints are worth using; this keeps your own judgement next to each model.

Features

• Rate each model from 1 to 3 stars with one click on the row under the selector.
• Comment each model: best sampler and scheduler, steps and CFG, resolution, known problems... (single line, up to 300 characters).
• Everything visible at a glance: stars are shown in front of the model name in the dropdown and in the selected value, so you can scan the list and spot your favourites.
• Drop-in replacement for the standard UNET loader: outputs MODEL and has the same unet_name and weight_dtype inputs.
• Your data stays local: plain JSON files, nothing is sent anywhere.

Usage

1. Pick a model in unet_name.
2. Click star 1, 2 or 3 to set the rating (clicking the last active star lowers it, down to none).
3. Click the comment field on the left, type your note and confirm.

Ratings and comments belong to the model, so they are shared by every MI Unet Rate node and every workflow.

---

MI Embedding

Pick textual-inversion embeddings from a dropdown, set the strength of each one and switch everything on or off, without typing embedding:name by hand.

Features

• Dropdown of your embeddings: every file ComfyUI sees in the embeddings folder (including folders from extra_model_paths.yaml).
• Strength per embedding: from -10 to 10, default 1.0.
• Enable / disable toggle: when disabled, the incoming text passes through untouched, so you can compare with and without embeddings without rewiring.
• Several embeddings in one node: + / - buttons (up to 8 rows).
• Before / after text: choose where the embeddings go relative to the incoming prompt.
• Subgraph friendly.

Usage: optionally connect a STRING prompt to text, pick an embedding and set its strength, then connect the text output to a CLIPTextEncode.

Rows	Text in	Position	Output
epiCPhoto.pt, 1.0	1girl	before text	embedding:epiCPhoto, 1girl
epiCPhoto.pt, 1.3	1girl	after text	1girl, (embedding:epiCPhoto:1.3)
epiCPhoto.pt, 0.8	none	before text	(embedding:epiCPhoto:0.8)
epiCPhoto.pt, 1.3	1girl	disabled	1girl

Good to know: embeddings must match your model family (SD1.5, SDXL...). The list is read at startup.

---

MI Compare 3 Images

Compare up to three images inside the node itself.

Features

• Three optional IMAGE inputs: image1, image2, image3 (a muted or unconnected branch never raises an error).
• Show 1 / Show 2 / Show 3 buttons switch the displayed image.
• Wipe 1-2 / Wipe 1-3 / Wipe 2-3 buttons compare a pair: hover the viewer and a bar follows the cursor, revealing one image or the other.
• Save 1 / Save 2 / Save 3 buttons save the matching image to the output folder as MM-dd-HH-mm-ss-imageN.png, and play the same confirmation sound as MI Save Image (c1.wav) when it is saved.
• Auto S. switch beside each Save button (orange = on): that image is saved to output automatically after every run, no click needed. Off by default; each switch is stored with the node.
• The resolution of each image is shown under its Save button.
• Sound when the images are ready: pick a WAV in the sound dropdown (same sounds as MI Reroute Sound, (none) keeps it silent). It plays in your browser once all the connected images have loaded and are ready to view, not when a saved workflow is reopened. Default: ding.
• The viewer fills the node and scales when you resize it.
• Images and the selected view are remembered when you switch workflow tabs or reload.
• Save Workflow switch (orange = on): embeds the workflow in the saved PNG so it can be dragged back into ComfyUI (skipped if ComfyUI runs with --disable-metadata).

Usage: connect up to three images, run the workflow, pick with Show N, compare with Wipe A-B, save with Save N. Previews live in ComfyUI's temp folder, so they disappear after a restart until the workflow runs again.

Inspired by the Image Compare Pixaroma node from ComfyUI-Pixaroma (MIT). This is an independent, simplified reimplementation.

---

MI Save Image

Receives one or several images and saves them to ComfyUI's temp folder with a preview. Nothing reaches your output folder until you decide: the SAVE button copies the visible image (or all images in grid view) there. It also passes the images through, so you can keep chaining nodes.

Features

• Preview first, save later: no more output folder full of discarded results.
• SAVE button on the node. A short confirmation sound (c1.wav) plays when the image has been copied to output.
• Date-based file names: filename_prefix defaults to %date:MM-dd-hh-mm-ss% (tokens yyyy, yy, MM, dd, hh, mm, ss). A subfolder can be included, e.g. portraits/%date:yyyy-MM-dd%.
• Never overwrites: if a name exists, _1, _2... is appended.
• Sound when an image is ready: pick a WAV in the sound dropdown (same sounds as MI Reroute Sound, (none) keeps it silent) and it plays in your browser each time an image arrives at the node. Default: ding. If the node is cached and does not run again, it does not play.
• Save Workflow switch (orange = on, same as in MI Compare 3 Images), next to the SAVE button. On: the workflow and prompt stay embedded in the saved PNG, so it can be dragged back into ComfyUI. Off: the image is saved clean, without them. The choice is stored with the node.

---

MI Reroute Sound

A minimal reroute: whatever you connect passes through unchanged, and when the flow goes through it the chosen WAV sound plays. Handy to know that a long generation has reached a point, without watching the screen.

Features

• Accepts any type (image, latent, model, text...) and returns it untouched, like a normal reroute.
• Tiny by design: it starts collapsed, just three orange notes (♪♪♪) with the input and output dots.
• Sound picker: expand the node (click the dot at the left of the title) to choose a .wav from the sounds/ folder. Picking one plays it as a preview. (none) keeps it silent.
• Includes seven small sounds (ding, chime, beep, done, soft, low, c1). Drop your own .wav files into sounds/ and restart ComfyUI to see them in the list.

Usage: put it on any link (for example after the sampler or after MI Save Image), choose a sound and run the workflow. The sound plays in your browser when the node is executed. If a node is cached and does not run again, it does not play.

---

MI Bounding Boxes

A box editor drawn directly on the node's canvas, so it works in the classic LiteGraph frontend (no Vue widget). It produces the same data as the core Create Bounding Boxes node, so you can wire it to MOSS-Image Layout Prompt, Build JSON Prompt (Ideogram) or the Flux layout prompt.

Usage

• Drag on an empty part of the canvas to draw a box; drag a box to move it; drag a corner handle of the selected box to resize it.
• Each box gets a row under the canvas with its number (in the box colour), a short text field (text to render; leave it empty for a plain object) and a description field. Click a field and type; long text wraps and the node grows.
• Buttons: Back / Front (change the order: first box = background, last = foreground) and Delete.
• Boxes are numbered 01, 02... in the canvas and in the preview; MOSS uses these numbers in its character_boxes field.
• Optional background image input: used as the canvas background and its size replaces width / height (shown after the first run).

Outputs: preview (IMAGE), bboxes (BOUNDING_BOX, pixels), elements (ARRAY, bbox = [ymin, xmin, ymax, xmax] on a 0-1000 grid), width, height, character_boxes (STRING such as 1, 3: the boxes whose checkbox at the right of the row is ticked; wire it to the character_boxes field of MOSS-Image Layout Prompt).

---

Where the data is stored

Personal data is kept in the data/ folder of this package and is ignored by git:

File	Content
triggers.json	LoRA name to trigger words (MI Lora)
hashes.json	Cached SHA256 of your LoRA files, for the Civitai lookup (MI Lora)
ratings.json	Model name to rating, 1-3 (MI Unet Rate)
notes.json	Model name to comment (MI Unet Rate)

Back them up if you want to keep them. Nothing leaves your computer except the LoRA hash you send to Civitai when you press the search line.

Repository layout

MI_nodes/
├── __init__.py        # registers all nodes
├── nodes/             # one Python module per node
├── js/                # frontend extensions
├── sounds/            # .wav files for MI Reroute Sound
├── images/            # banner and screenshots
└── data/              # your ratings, notes, trigger words and hashes (not in git)

License

MIT © 2026 Thinkingbells